Internship at Lennar — Applied AI Engineering

OverviewRole & scope

Architected scalable FastAPI microservices with object-oriented design supporting multimodal AI workflows — processing 1000+ images with hybrid search (semantic + keyword + reranking) in 240 seconds via async batch processing and efficient API orchestration.
Built high-performance vector-database infrastructure using FAISS with O(1) metadata lookup and hybrid retrieval, enabling real-time semantic search across construction-domain knowledge with sub-second query response.
Engineered an end-to-end NLP pipeline turning construction conversations into structured reports — few-shot classification, K-means / HDBSCAN clustering, and a RAG framework with LangChain — reaching 20k+ text embeddings in under 300 seconds.
Optimized ML system performance through asyncio and concurrent OpenAI API calls, delivering measurable precision/recall gains on production-scale noisy text while keeping the system reliable.

Automated quality-control extraction from construction communications: identify a QC issue, extract it, enrich with details, and output an issue form.
Image parsing: keyword labelling for each site image.
Document-image parsing: extract structured data from each document image.
Image search queryable in natural language.

Input

A single target message, supported by previous messages and external data pulled from the existing database.

Output

The extracted issue plus an evaluation report — produced through data preprocessing, classification & clustering, extraction, and enrichment.

Input: parsed images plus a user query. Output: the top-k most relevant images — searchable in plain language, no manual tagging required.

Module

Extracts visual features and metadata from site images.

Module

Analyzes construction documents, QC reports, and forms.

Module

Resolves natural-language queries to the most relevant images.